Abstract

Residual network is a deep learning model widely used in various applications of computer vision, such as image processing, semantic classification and video processing. The traditional residual network is susceptible to the interference induced by the adversarial example attack algorithms, which directly affects the security of practical applications. In this work, we propose a new residual neural network based on a global convolutional denoising model. The global convolution is fused with the denoising technique, and the network structure of the global convolution denoising module is redesigned so that the denoising module can be trained end-to-end with the residual neural network. In addition, the gradient information of the network is concealed to enable the neurons to respond to the pixels that are more meaningful to human vision, thus improving the robustness of the network for adversarial example attack algorithms.

中文翻译：

---

基于全局卷积降噪模型的对抗样本生成方法

摘要

残差网络是一种广泛应用于计算机视觉各种应用的深度学习模型，例如图像处理、语义分类和视频处理。传统残差网络容易受到对抗性实例攻击算法的干扰，直接影响实际应用的安全性。在这项工作中，我们提出了一种基于全局卷积去噪模型的新残差神经网络。将全局卷积与去噪技术融合，重新设计全局卷积去噪模块的网络结构，使得去噪模块可以与残差神经网络进行端到端的训练。此外，隐藏网络的梯度信息，使神经元能够响应对人类视觉更有意义的像素，